ikernInt

ikernInt integrates supervised and unsupervised analyses of spatio-temporal metagenomic NGS datasets using compositional kernel methods to handle compositionality and enable phenotype prediction, ecological dissimilarity-based clustering, and retrieval of microbial signatures via taxa importances.

Key Features:

• Unified framework: Integrates supervised learning (predicting phenotypes from taxonomic abundances) and unsupervised methods (clustering and visualization based on ecological dissimilarities) within a kernel-based framework that evaluates taxa importances.
• Compositional kernels: Implements two compositional kernels—Aitchison-RBF and compositional linear—designed to handle the compositional nature of NGS data.
• Beta-dissimilarity transformation: Transforms non-compositional beta-dissimilarity measures into kernel functions suitable for kernel-based analyses.
• Spatial integration via multiple kernel learning: Employs multiple kernel learning to incorporate spatial data into analyses.
• Temporal kernels for longitudinal data: Provides specific kernels tailored to evaluate temporal variations in longitudinal datasets.
• Taxa importance retrieval: Facilitates retrieval of microbial signatures by evaluating taxa importances within the kernel framework.

Scientific Applications:

• Microbiome studies: Enables integrated analysis of spatial and temporal dynamics in metagenomic/microbiome research.
• Phenotype prediction: Supports prediction of phenotypes from taxonomic abundance profiles.
• Visualization and clustering of ecological dissimilarities: Converts ecological dissimilarities into kernel functions for visualization and clustering purposes.

Methodology:

Defines compositional kernels (Aitchison-RBF and compositional linear) that transform beta-dissimilarity measures into kernel functions, applies multiple kernel learning to integrate spatial data, uses specific temporal kernels for longitudinal analysis, and evaluates taxa importances to retrieve microbial signatures.